Does AI Expand Human Capability—or Just Output?

We already have good evidence for short-term productivity

Controlled studies have found meaningful gains on bounded professional tasks. In a randomized experiment with 453 college-educated professionals, access to ChatGPT reduced average completion time by 40% and raised assessed quality by 18% on the studied writing tasks. The current NBER record for Generative AI at Work reports a 14% average productivity increase across 5,179 customer-support agents, with a 34% improvement among novice and lower-skilled workers. In the BCG jagged-frontier experiment, AI users completed 12.2% more within-frontier tasks and worked 25.1% faster, but were 19% less likely to be correct on a selected task outside the model's frontier. Sources: Science · NBER · BCG / Organization Science

Those results are important. They also answer a narrower question than the one we care about most.

Short-term performance does not tell us what persists

If AI helps someone produce a strong answer today, what happens tomorrow when the AI is removed?

Did the person learn the underlying skill? Did they gain a better mental model? Did the saved time move into higher-value work? Or did the workflow become dependent on the assistant?

Most short-term productivity studies are not designed to answer those questions.

The same tool can help one task and hurt another

The jagged-frontier research is especially important here. AI improved performance on tasks inside its frontier and reduced correctness on a task outside it.

That means human capability expansion is not simply “give people more AI.”

People also need better ways to understand when assistance is helping, when it is masking a gap, and when it is pulling them outside the system's reliable range.

What we would measure over time

We would measure more than output volume.

The research question is what becomes possible

GO Beyond does not need a story about one person replacing an entire company.

The research question is what new objectives become reachable when expertise, tools, and execution support become more accessible.

That may mean a founder can test an idea faster, a small team can analyze a problem that once required a specialist, or an individual can communicate, build, or learn at a level that was previously impractical.

The human remains responsible for purpose, values, and consequential judgment. The technology should expand the frontier around them.

A longitudinal test

We want longitudinal evidence that separates immediate assisted performance from durable human learning.

That means measuring people both with and without AI over time, tracking skill transfer, time reallocation, dependence, subgroup effects, and whether the gains persist when the task changes.

Until that research exists, “AI expands human capability” should remain a product thesis and research question—not an empirical claim that the problem has already been solved.

What would count as an answer

AI can increase immediate output on some tasks.

That does not by itself show that people become more capable over time.

A useful answer would require evidence that the gains persist, transfer to new work, or expand what people can do when the immediate assistance is no longer present.